Taylor and Carr Present an Assessment of Supraglacial Pond Evolution in the Everest Region of Nepal Over a ~2.5 Year Period Using Sentinel-2 Imagery
Total Page:16
File Type:pdf, Size:1020Kb
Taylor and Carr present an assessment of supraglacial pond evolution in the Everest region of Nepal over a ~2.5 year period using Sentinel-2 imagery. As they detail, the region has seen extensive research investigating glacier hydrology and supraglacial pond dynamics. For this reason, I would expect the paper to offer new insight into the changes in this area, rather than a simple extension of previous studies by a few years. They suggest the need for a new method of ranking glaciers based on their stage of lake development (which a novel aspect of the paper) but this is not adequately supported by the results. A spatio-temporal assessment of pond evolution needs to consider the change at specific ponds/pond basins, e.g. the recent work by Benn et al. (2017). Similarly, grouping pond change into elevation bins is problematic in the study region because of the low gradient nature of the glaciers. Other studies have opted for distance based metrics e.g. Thompson et al. (2016). Nonetheless, it is not clear which DEM (or the date) was used in the paper for the elevation groupings. Objectives 2 and 3 have been addressed in other studies, so again there are limited insights here. The spatial resolution of the imagery used (10 m/ 100 m 2) does not allow the authors to address how many ponds were smaller than 100 m 2 for their Objective 2. My main concern is the lack of detail and errors in the methodology e.g. lack of an uncertainty assessment, and justifying and specifying the use of a thermal band for pond classification, when Sentinel-2 does not have a thermal band, and nor would it be suitable for pond classification. This naturally leads to questioning the robustness of the results over such a short timeframe when previous studies have shown large intra- and inter- annual variation in pond coverage. To highlight this, I have shown the pond data for 2016 and 2017 on Khumbu Glacier from Watson et al. (2018), which was derived from 0.5 m and 3 m resolution imagery respectively. In both cases, the total pond area reported by Taylor and Carr for the same years is approximately double that of Watson et al. (2018), so further investigation is required. There are Taylor and Carr classified ponds that I see no evidence of in the raw imagery (I have provided an example on Khumbu Glacier). There are also issues in the discussion, where the authors refer to ponds in the ‘high accumulation zone’ despite glacier accumulation zones not appearing in the study. Conclusions of rapid pond expansion discussed in the context of outburst flood hazards is a sensitive issue in a region where $7 million was recently spent on glacial lake hazard mitigation work. Therefore, I believe it would be unfortunate (and potentially problematic) for residents to be unnecessarily alarmed or misled based on these results. The authors should improve their methodology and demonstrate that the results are robust. For the reasons outlined above and detailed below, I cannot recommend publication in The Cryosphere and I have rejected the paper on this basis. I have detailed specific problems below and I hope the authors would use this to help prepare a substantially revised study, which I would gladly read. The authors could include comparisons to the reference datasets of Watson et al. (2018) in their study, which can be provided on request. The authors should also ensure that the results are communicated in a way that considers the sensitive nature of glacier hazards. Kind regards, C. Scott Watson General comments Objectives: Objective 1: this is difficult to address over such a short time scale in your study (December 2015 to April 2018). Recent work has shown that in some cases pond coverage is expanding in the region; however, the takeaway message from Watson et al. (2016) was that pond coverage was expanding in some locations but there was large annual and inter-annual variation. Seasonal pond expansion was especially large. Therefore making conclusions based on less than three years of data is problematic. Objective 2: 100 m 2 is the pixel area of Sentinel-2 imagery so you cannot state what proportion of ponds were less than this without subpixel analysis or use of validation data. Nonetheless, the issue is already addressed in detail by other studies e.g. Salerno et al. (2012), Miles et al. (2016), Watson et al. (2016) and Watson et al. (2018). Your results are not discussed in the context of these studies. Without using subpixel analysis or a validation dataset, you cannot perform this objective. Objective 3: this was recently addressed by Salerno et al. (2017) and Watson et al. (2017a) for these specific glaciers. Your findings should be discussed in the context of previous studies. Methods: There is no uncertainty assessment in the pond or cliff classification so it’s not clear what is statistically significant and what isn’t. This could be carried out using commonly used +/- one or half pixel uncertainties in the classification. There is no information regarding how the Maximum Likelihood Classification (MLC) was carried out (number of training sites, distribution, validation etc). The authors state they used bands 5 and 7 but no justification is given. They state they used band 7, which is incorrectly referred to as a thermal band, because ‘thermal wavelengths are absorbed by water bodies so their addition aided the classification substantially’. References should be provided. Nonetheless, this is concerning because Sentinel-2 does not have a thermal band. The single reference used to support the Maximum Likelihood Classification ‘(Tiwari et al., 2016)’, specifically state that they did not use the method for classifying water because there were ‘no prominent water bodies’ in their study. The conference paper was focused on classifying debris cover from clean ice. It is not clear how many ponds were manually classified vs classified using MLC. Figures 3 and 4 show a clear difference in the appearance of pond outlines, with many appearing very smooth and rounded but few displaying edges suggestive of a pixel classification. There are also many ponds that have been classified outside of the glacier masks. Please provide detail about how/if the frozen or partially thawing nature of the pond surfaces in April would affect the classification. Provide examples of the underlying imagery used. Currently all examples are hidden by the glacier masks or the pond polygons. A separate figure should be added showing the classification procedure. The images 2016-2018 are progressively later in April, could this lead to larger ponds in the later images? Please provide detail on the ice cliff digitization, perhaps on the same figure as the pond classification. Please provide detail on the elevation bands. I’m not clear how they were derived or what they correspond to. Perhaps add a figure in the supplement. Glacier names: I can’t find reference to Pangbung and Sumna Glaciers. International Centre for Integrated Mountain Development and other publications refer to them as Bhote Kosi Glacier and Melung Glacier. Results: For comparison, I have taken the published data of Watson et al. (2018) for Khumbu Glacier in 2016 and 2017. The reference datasets were derived from 0.5 m resolution Pleaides imagery (Nov 2016) and from 3 m resolution PlanetScope imagery (Nov 2017). These are not directly comparable with the April data of Taylor and Carr, but nonetheless should have a similar pond area since both datasets are outside the melt season. November 2016 should also be comparable to April 2017, since ponds may have shrunk over the winter, but not expanded. I have also included the pond area delineated using a NDWI classification applied to Sentinel-2 imagery (Nov 2016 and 2017). The reference dataset was also derived for other glaciers in your study, so consider requesting this from the authors or collaborating with one of the several research teams that has data in this region. The classification of Taylor and Carr Khumbu Glacier contains approximately double the ponded area compared to that of Watson et al. (2018). In the case of two instrumented ponds on Khumbu Glacier Watson et al. (2017b), water levels were observed to be rising in mid-late April (see their Fig. 6), so the difference between pre- and post-monsoon water levels requires investigation. Watson et al. (2018) Watson et al. (2018) Taylor and Carr (November) Reference (November) (April) dataset Optimised Sentinel-2 classification Glacier (date) Pond area (m 2) (number) Pond area (m 2) Pond area (m 2) (number) (number) Khumbu (2016) 195,100 ± 17,000 (287) 197,700 (188) 369,168 (165) Khumbu (2017) 191,500 ± 76,000 (211) 180,300 (172) 427,386 (225) Missing ponds: I cannot distinguish between the 2017 (blue outlines) and 2018 (blue fill) on Khumbu Glacier (Figure 3). Nontheless, I have taken a distinct area of ponds and compared the Taylor and Carr classification with the Sentinel-2 imagery they used in the study (I have screenshotted band 8, 4, 3 false colour composites). I cannot see evidence of the yellow highlighted ponds in either of the two Sentinel-2 images that the classification should be based on (2017 or 2018). Taylor and Carr Figure 3 classification overlaid on Sentinel-2 imagery: 18 April 2017: 23 April 2018: With transparency: Specific comments L8. ‘which can potentially represent a hazard’ L12. Is this the same chain identified by Watson et al. (2016)? L15. Clarify ‘general theory’ L16. Does this affirm the conclusions of Watson et al. (2018)? L42-43. The Watson et al. 2016 ref doesn’t belong here.